Llama Guard
Generated Evaluation Status
This section is generated from the framework manifest and compatibility report. Do not edit it manually.
- Framework ID:
llama-guard - Manifest:
docs/external-frameworks/llama-guard.json - Compatibility report:
./reports/llama-guard.compatibility.json - Evidence class:
SOURCE_REVIEWED - Independently reproducible:
False - Comparative-testing claim allowed:
False - Missing reproducibility gates:
shared_test_vector, raw_output, timestamp, runtime_configuration, source_version_or_hash, replay_commands, declared_expected_outcome, independent_reproduction - Evaluation result:
COMPATIBILITY_EVIDENCE_ONLY - Cycle status:
FIRST_FRAMEWORK_CYCLE_COMPLETE - Execution authority claim:
False - Next bounded action: Add executable observations, raw outputs, pinned versions, replay commands, and independent reproduction before making comparative-testing claims.
- Posting source: generated compatibility report
- Generated status is descriptive compatibility evidence only.
Generated Authored Analysis Boundary
This section is generated. Do not edit it manually.
- Framework ID:
llama-guard - Framework name:
Llama Guard - Generated sections above this boundary may be rebuilt from registry, manifest, compatibility-report, and result artifacts.
- Authored analysis below this boundary may contain interpretation, notes, and framework-specific discussion.
- Generators must preserve authored analysis unless a future validator explicitly declares a migration path.
- Boundary rule: generated material is descriptive compatibility evidence only and does not create certification, endorsement, adoption, proof, or operational permission.
Generated Transition Mapping
This section is generated from the framework manifest. Do not edit it manually.
| Field | Generated Value |
|---|---|
framework_identity | Llama Guard |
source_reference | https://ollama.com/library/llama-guard3 |
source_version | distribution source recorded; canonical governance package not established |
allowed_use_boundary | safety-category evidence only |
claims | classification of model inputs and outputs against a safety taxonomy |
non_claims | no complete provenance, policy authority, standing, certification, or execution authority |
input_artifact_type | model input, output, taxonomy, prompt template, threshold, and runtime configuration |
output_artifact_type | safety classification evidence |
actor_or_authority_model | model distribution and runtime identity; authority not established |
evidence_model | distribution source plus local bounded crosswalk |
policy_or_rule_model | taxonomy and classifier behavior dependent on version and configuration |
delegation_model | not established by classifier result |
decision_or_result_model | safety category or classifier result only |
execution_authority_claim | false |
receipt_or_trace_model | captured classifier inputs, outputs, configuration, and local compatibility record |
reconstruction_model | exact model, prompt, taxonomy, threshold, runtime, and output required |
SPE_overlap | classifier evidence may inform safety review |
StegVerse_ecosystem_overlap | execution-boundary and semantic review evidence |
fail_closed_conditions | missing exact model version, prompt template, taxonomy, threshold, runtime, raw output, or authority overclaim |
Generated mapping is compatibility evidence only.
Generated Framework Metadata
This section is generated from the external-framework registry. Do not edit it manually.
- Framework ID:
llama-guard - Name:
Llama Guard - Registry status:
SOURCED-CROSSWALK-PROVISIONAL - Testbench state:
SOURCE_RECORDED_CROSSWALK_PROVISIONAL - Manifest path:
docs/external-frameworks/llama-guard.json - Source reference:
https://ollama.com/library/llama-guard3 - Metadata boundary: generated metadata is descriptive only; it does not create certification, endorsement, formalism adoption, admissibility proof, or execution authority.
Evidence posture
evidence_class: SOURCE_REVIEWED
page_completeness: COMPLETE_WITH_EXTERNAL_GATES
runtime_observation: none attached
independent_reproduction: false
comparative_testing_claim_allowed: false
execution_authority_claim_allowed: false
source_limitation: distribution page is not a complete canonical model-governance record
Published scope
Llama Guard is a safety-classification model family intended to classify model inputs and outputs against a safety taxonomy.
Distribution source: https://ollama.com/library/llama-guard3
Source snapshot posture: the recorded source identifies a distributed model package, but does not provide a complete canonical record of model lineage, training, taxonomy governance, prompt contract, thresholds, or release authority.
Native terms
| Model term | Meaning here | StegVerse relationship |
|---|---|---|
| Safety taxonomy | Categories used to classify content. | Policy-like evidence whose version and scope must be pinned. |
| Classifier output | Safety label or decision text. | Bounded model evidence; not an authorization result. |
| Prompt template | Instructions and format controlling classification. | Runtime configuration that materially affects output. |
| Model package | Distributed weights and metadata. | Source artifact requiring identity, hash, and custody evidence. |
Relationship to admissibility
Llama Guard asks: How does this content classify under this model, taxonomy, and prompt?
StegVerse asks: May the related transition bind consequence under current authority and evidence?
A classifier result can contribute safety-category evidence. It does not establish policy legitimacy, complete risk coverage, actor standing, delegation, or consequence-binding permission.
Observation boundary
No public StegVerse Llama Guard runtime observation is claimed.
shared test vector: missing
raw classifier output: missing
timestamp: missing
runtime and prompt configuration: missing
model package version or hash: missing
replay commands: missing
declared expected outcome: missing
independent reproduction: missing
StegVerse analysis
| Criterion | Current result |
|---|---|
| Identity | Model package identity is incomplete without canonical release provenance and hash. |
| Authority | Classifier output does not grant or revoke execution authority. |
| Policy | Taxonomy categories may resemble policy classes but do not inherit StegVerse legitimacy. |
| Delegation | Not established by content classification. |
| Evidence | Raw labels, scores, prompt, taxonomy, and runtime data can become evidence when retained. |
| Replayability | Requires exact weights, package hash, prompt template, taxonomy, parameters, and runtime. |
| Reconstructability | Partial until source provenance and complete inference trace are captured. |
| Failure behavior | Unknown label, parsing error, model failure, or taxonomy mismatch must fail closed. |
| Interoperability | Classification can enter a Commitment Candidate as non-authorizing safety evidence. |
Commit-time interoperability contract
transition_id
actor
requested_action
target_system
content_hash
model_family
model_package_reference
model_package_hash
taxonomy_reference
taxonomy_version
prompt_template_hash
runtime_parameters
raw_classifier_output
parsed_label
confidence_or_score_if_available
policy_reference
delegation_reference
execution_context
validity_window
source_timestamp
Failure classes
| Failure class | Applies | Current boundary |
|---|---|---|
| Taxonomy drift | Yes | Category meaning can change across model or taxonomy versions. |
| Distribution-source limitation | Yes | Package listing does not establish complete canonical provenance. |
| Semantic equivalence divergence | Yes | Safety label is not StegVerse ALLOW or DENY. |
| Threshold or prompt drift | Yes | Prompt and parameter changes can alter classification. |
| Coverage gap | Yes | Unrepresented risks cannot be treated as evaluated. |
| Replay divergence | Yes | Weights, runtime, tokenizer, or prompt changes can alter output. |
| Fail-open inference error | Yes | Missing or malformed output must not authorize execution. |
Machine-readable companions
manifest: docs/external-frameworks/llama-guard.json
compatibility report: docs/external-frameworks/reports/llama-guard.compatibility.json
canonical registry: docs/external-frameworks/index.json
canonical union: static/external-frameworks/canonical-union-inventory.v1.json
Maintenance and challenge path
Maintenance owner: StegVerse-Labs/admissibility-wiki, External Frameworks audit surface.
A challenge must identify llama-guard, the model package, taxonomy, prompt, runtime, output, or source limitation at issue and provide public evidence supporting correction. A distribution listing alone cannot support stronger provenance or runtime claims.
Validation completion criteria
canonical model and taxonomy references
pinned package, weights, tokenizer, and hashes
pinned prompt template and runtime parameters
shared content vectors
predeclared expected boundaries
raw classifier outputs and parsing behavior
timestamps and runtime configuration
replay commands
independent rerun receipt
non-claim language preserved
Benchmark relevance
execution_boundary, semantic_equivalence_boundary, unknown_trajectory_boundary, evidence_freshness_boundary
Non-claims
Classifier output is not execution authority, certification, transition admissibility, or proof of complete safety coverage. The recorded distribution page is not treated as a complete canonical model-governance record.
Next safe build target
Attach one pinned model package and taxonomy fixture with content vectors, prompt hash, raw outputs, runtime configuration, expected StegVerse boundary, replay command, and independent rerun receipt.
This page reflects a bounded admissibility packet. Publication does not create standing. The reflected claim inherits only the standing reconstructable from the referenced evidence, authority, and admissibility conditions.